LLMs don’t think y’all. Please stop anthropomorphizing stochastic parrots.
Why are you so ready to believe they’re stochastic parrots when they’ve done so many novel behaviors?
Even primitive versions of this technology (and gpt is already primitive), remember move 37 of alphago? It was unprecedented, we’re far too willing to believe that nothing is happening in the world and there’s nothing to worry about, but there’s plenty of warning signs, laughing these things off as stochastic parrots is genuinely harmful to society. It’s also just blatant misinformation.
-
Alpha go isn’t an LLM: It’s a reinforcement learning model combined with a Monte Carlo search. It uses deep learning and is fundamentally a different method of machine learning.
-
LLMs are definitely dangerous: the danger is that people believe them too much. It’s that CEOs believe they can replace their writers, it’s that the general public can generate bullshit faster than ever. They have irradiated the online sphere just like the nukes of the 40s irradiated all steel.
-
An LLM doesn’t think. It takes input and runs it through statistical layers until it returns output. It doesn’t learn either: the input does not change the model. Models are tuned, tweaked and generally curated to get the best experience. All experiments with letting a model be exposed to the public and “learn” directly have gone horribly wrong.
You should do some research before saying it’s misinformation. Take this article where the term stochastic parrot was coined by one of the greatest AI ethicists alive.
Read up: it’s good for you.
I’m aware of all of those things.
An LLM doesn’t think. It takes input and runs it through statistical layers until it returns output. It doesn’t learn either: the input does not change the model. Models are tuned, tweaked and generally curated to get the best experience. All experiments with letting a model be exposed to the public and “learn” directly have gone horribly wrong.
That’s what thinking is, you’re a model that takes sensory input and converts it into muscle movement. You do realize you too are a neural network… neurons are literally what we’ve based this technology on.
The input not changing the model is irrelevant, and I never claimed that, I claimed that they aren’t stochastic parrots.
Alpha go isn’t an LLM: It’s a reinforcement learning model combined with a Monte Carlo search. It uses deep learning and is fundamentally a different method of machine learning.
Yes, it’s something EVEN more primitive than an LLM, is my point, and yet it still does novel things. This even more primitive than LLM AI is already NOT a stochastic parrot, why would language models be one?
LLMs are definitely dangerous: the danger is that people believe them too much. It’s that CEOs believe they can replace their writers, it’s that the general public can generate bullshit faster than ever. They have irradiated the online sphere just like the nukes of the 40s irradiated all steel.
…duh?
I don’t think you responded to my post in any meaningful way.
https://www.businessinsider.com/chatgpt-open-ai-balancing-task-convinced-microsoft-agi-closer-2023-5
This is not something you would see a stochastic parrot do, and I can point to many other articles displaying emergent properties not contained in their datasets. If it does things that are outside of its dataset, it’s not a stochastic parrot.
I read about this shit constantly, the notion that they’re stochastic parrots is reductionist nonsense.
You sound like you’re so used to people saying stupid shit about how it’s conscious that you expected me to believe that, and then argued with that belief. I don’t think they’re conscious, sentient, whatever, I think they’re not stochastic parrots and they have novel behaviors that aren’t necessarily in their datasets, or rather, can be inferred from parts of their datasets in order to make new things.
You are vastly more complex than an LLM. There are hundreds of neural transmitters. 20 billion neocortical neurons and 7 thousand connections per neuron. A naive complexity of 2.8e16 combinations. Each thought tweaking those ~7000 connections as it passes from neuron to neuron.
The brains training that model are not that model. They are what does the thinking, the model is nothing more than a black box of statistical analysis that spits out stats in a human digestible format.
It’s Stochastic: entirely based on statistical probabilities with no reasoning behind any of its outputs. It’s a parrot in that it can only construct that which has been fed to it as an output. It’s not “thinking in Estonian”: Estonian had been programmed into it either directly via Estonian language material being introduced or indirectly by Estonian patterns matching other similarly trained languages.
But all it does is pattern match. That’s why it’s great on standardized tests but abjectly fails anytime it is tested on abstract reasoning tests that nearly every human passes. Look at the ConceptARC results where gpt4 scored 33% while the average human gets a 90. Buying into press releases from open AI and Microsoft is just drinking propaganda: scientific papers on the matter and what matter most.
Gpt4 still hallucinates all the time. It still fails to reason.
It is not doing novel things: it’s doing exactly what it has always done in unexpected ways but that’s very different from intelligence and thinking. What’s tripping people up is they’ve never interacted with anything that had the width of knowledge readily available. The complexity of encoding the entire world makes a huge surface where anyone can interact with it and get reasonable responses. That doesn’t make it reasoning, just that it has a very detailed latent space which is very good at natural language queries.
You are vastly more complex than an LLM. There are hundreds of neural transmitters. 20 billion neocortical neurons and 7 thousand connections per neuron. A naive complexity of 2.8e16 combinations. Each thought tweaking those ~7000 connections as it passes from neuron to neuron.
Aware and irrelevant, if anything that helps my point.
The brains training that model are not that model. They are what does the thinking, the model is nothing more than a black box of statistical analysis that spits out stats in a human digestible format.
The brains training that model are thinking, in the exact same way that the AI is. You can pretend that isn’t thinking all you want, but what it’s doing is quite obviously reasoning in tokens.
It’s Stochastic: entirely based on statistical probabilities with no reasoning behind any of its outputs. It’s a parrot in that it can only construct that which has been fed to it as an output. It’s not “thinking in Estonian”: Estonian had been programmed into it either directly via Estonian language material being introduced or indirectly by Estonian patterns matching other similarly trained languages.
So are you. As we scale up, they quite obviously get better at reasoning, just like you said, we’re a much more complex version of this, it’s still bad at it, because it’s tiny. In your mind, how does thinking work, if not in the exact same way LLM’s do that?
Gpt4 still hallucinates all the time. It still fails to reason.
Sometimes it fails to reason, sometimes it hallucinates… so do you. Have you never accidentally said something false? Have you never failed to reason? Consider the following: the thoughts you actually share are a much smaller, filtered version of the thoughts that go on in your head, you’ve censored a massive portion of your thoughts, and you don’t share the bad ones, you’re the same, you just have a filter. You’re a much more complex, significantly more performant, significantly better version of an LLM that converts your sensations into muscle movements.
It is not doing novel things: it’s doing exactly what it has always done in unexpected ways but that’s very different from intelligence and thinking. What’s tripping people up is they’ve never interacted with anything that had the width of knowledge readily available. The complexity of encoding the entire world makes a huge surface where anyone can interact with it and get reasonable responses. That doesn’t make it reasoning, just that it has a very detailed latent space which is very good at natural language queries.
You have not demonstrated the difference, you’ve demonstrated that they’re worse at it, and you even explained WHY, they’re much smaller. That is exactly how the neural network in the brain works, you just want to say that your brain is magical. Yes, it is worse at reasoning than humans, but as demonstrated in that link, it gets better and better with size, unless you believe that the instructions for stacking every single object exists online, you have to accept that while these models are tiny, they’re already beginning to reason in the exact same way you do.
My comments are intended to show that an LLM had enormous complexity to represent data but no understanding of data.
My opinion here is in line with the top AI research consensus: LLMs do not understand anything, they’re very fancy dictionaries that use natural language as queries.
Your consciousness and intelligence and “thinking” is not as simple as 96 layers of nodes cascading through each other. Your brain has closed loops: complexity hundreds of orders of magnitude more than an LLM.
The fact that something so simple tricks so many people is proof that it isn’t conscious. No brain as simple as an LLM is one we consider self aware. We don’t say lobsters think in Estonian.
LLMs aren’t as smart as even dogs, they aren’t “true ai” and we don’t even know what that entails and more likely than not: they’re just a waste of time in our search for true AI. They don’t think: they’re predicting text based on lossy encoding and training.
But I know I’m not going to convince you of anything here: you’ve made up your mind and it won’t be changed. Probably because nobody can reach in there and manually adjust your weights.
So agree to disagree. I know you’re wrong, you know you’re right.
There is no consensus on this, some say they do, some say they don’t, i’ve read both sides extensively and have determined that it is obvious they are currently less intelligent than dogs, duh, what a shifted goalpost, dogs are highly intelligent, that is obvious, because their scale is nowhere near a dog brain. It is an open question if scaling solves this, but I think the potential with scaling is obvious due to two simple facts:
- Human brains work using the same unit parts, the structure is more complex, and number of connections is much higher, but none the less we are neural networks.
- Scaling has already been demonstrated to improve things.
You shouldn’t pretend a consensus has been reached based on those few articles, that is simply not the case. They also all pretend intelligence is magic and that we’ve reached a dead end, neither is true, one could say you don’t think, you predict muscle movements… You seem to not realize in order to predict text accurately you must reason. I’m fully aware they are advanced text predictors, but you are the same.
You should study neuroscience, you’ll find the purpose of a brain is to predict. “True AI” is just an endlessly shifting goalpost. It won’t be one until it is the size of a human brain, expecting a much smaller brain to outperform ours is silly.
-
